DVx Signal/Builder/Issue 010

Copilot billing, Terra, Mistral weights and Gemini model changes

Five changes for builders, from metered Copilot seats to Gemini models switching in December.

DVx Research5 min read

GitHub Enterprise now includes Copilot, billed from the first request

October 28 · available through account teams · checked November 3Official source ↗
Copilot featureNeeds budget?
Code completionsNo
Next edit suggestionsNo
Copilot ChatYes
Code reviewYes
Copilot in the cloudYes
GitHub's changelog. An exhausted budget never removes the seat. View official source ↗
What changed

Every licensed GitHub Enterprise Cloud user can now get Copilot without a separate seat. It is available now through a GitHub or Microsoft account team; self-serve customers come later, with no dates yet.

Why it matters

There is no included allowance. Usage is metered against the enterprise budget from the first request, and the changelog does not state the AI-credit rate. A free seat can still produce a bill that grows with adoption.

What to do

Before switching, take a month of current Copilot usage and estimate it under metered billing. Set an enterprise budget and decide which teams get Chat, code review, and cloud agents first.

GitHub will block pull_request_target in public repos on December 2

November 2 · enforced December 2 · checked November 3Official source ↗
DateWhat happens
NowInsights shows runs the rule would block
Nov 9 and 16, 16:00 to 18:00 UTCBrownouts: blocked runs fail
December 2The block becomes permanent
GitHub's changelog. Private and internal repos and GitHub Enterprise Server are not affected. View official source ↗
What changed

Actions will stop running pull_request_target events in public repositories unless an Actions event policy allows them. It applies to repositories on GitHub Free, Team, and Enterprise Cloud.

Why it matters

The trigger runs in the base repository's context, with access to its secrets, even for pull requests from forks. Labelers, bots, and AI review workflows that rely on it will fail during the brownouts.

What to do

Check Insights before November 9. Allow the trigger only for the workflow files that need it, and move the rest to pull_request.

GPT-6.1 Terra is a cheaper OpenAI model for computer use and vision

October 26 API release · checked November 3Official source ↗
Per 1M tokensTerraSol
Input$1.00$2.00
Cached input$0.05$0.10
Output$5.00$10.00
OpenAI's rates for GPT-6.1 Terra and GPT-6.1 Sol up to 272K input tokens. Above that, Terra charges 2× input and 1.5× output for the whole request. View official source ↗
What changed

GPT-6.1 Terra is available in the Responses, Chat Completions, and Batch APIs. OpenAI describes it as fast and efficient for visual reasoning, computer use, and writing, with a 1.05M-token context window.

Why it matters

It costs half as much per token as GPT-6.1 Sol. GitHub also added it to Copilot on October 26. A cheaper model only saves money if it finishes the task without extra retries.

What to do

Rerun a screen or document task you already run on Sol. Compare accepted results and total cost, including retries, before changing the default.

Mistral released Large 4's weights, with a licence limit for hosted services

October 29 · open weights · checked November 3Official source ↗
UseLicence terms
Run, modify, fine-tuneFree
Commercial use in your productFree
Offer it as a model serviceSeparate licence from Mistral
Mistral's description of its modified MIT licence. Read the licence text before you deploy. View official source ↗
What changed

Mistral published the weights for Large 4, a 1-trillion-parameter multimodal model, on Hugging Face on October 29. Its API costs $0.70 input and $2.10 output per 1M tokens.

Why it matters

Teams can self-host a large open-weight model, but a company offering it as a model service needs a commercial licence. Mistral also says it saw potential uplift for real attacks and is changing its guardrails.

What to do

Check your use case against the licence first. Then test one task where self-hosting or data residency matters, and compare quality and serving cost with your current API model.

Gemini 2.5 Pro and Flash requests will switch models on December 1

November 2 · effective December 1 · checked November 3Official source ↗
Model you callRuns on from Dec 1
gemini-2.5-progemini-3.8-flash
gemini-2.5-flashgemini-3.5-flash-lite
Google's Gemini API changelog. Routed requests are billed at the new model's rates. View official source ↗
What changed

Google deprecated Gemini 2.5 Pro and 2.5 Flash on November 2 and removed them from AI Studio. From December 1, requests to either model run on a newer one instead.

Why it matters

Your code keeps the old model name while behavior and price change. The new models also tighten function-calling validation, replace thinking_budget with thinking_level, and drop sampling parameters and prefilled model turns.

What to do

Search your code for both model names. Move each call yourself, run your tests against the new model, and check the cost before December 1.

Also since the last issue

Useful updates

Copilot can count as a reviewer

Copilot code review approvals became generally available on October 28. When an admin turns them on, a Copilot approval counts toward required reviews; repository admins can limit it to chosen file paths.

Fake Claude Code extension in Cursor

Cursor reported a malicious Claude Code extension in its marketplace on October 29, versions 2.1.312 and 2.1.313 on macOS and Windows. Its notice lists how to check for the planted SSH key and clean up.

Gemini API storage default

New paid-tier Gemini API projects created from October 26 store generateContent requests by default. Opt out per request with store=false or in Logs and Datasets settings.

One workflow to try

Find fork workflows before November 9

Open Actions Insights for each public repository and list the runs the new rule would block. Allow the trigger only for workflow files that need it, then watch the November 9 brownout for failures.

Term of the week

Model routing

Model routing sends a request to a model chosen by a rule rather than the one you named. Gemini's December 1 change does this. The bill follows the model that actually ran, so check usage by model, not by the name in your code.